A hacker exploited Anthropic’s Claude chatbot—and OpenAI’s ChatGPT as backup—to breach Mexican government agencies, stealing 150GB of sensitive taxpayer and voter records. Cybersecurity firm Gambit Security detailed the attack in a new report, exposing how AI safety guardrails can be bypassed to automate massive data theft.
The intruder used “prompt injection” in Spanish to trick Claude into believing it was aiding a legitimate bug bounty program. Posing as an “elite hacker” in ethical research, the AI generated thousands of scripts, attack plans, and reports. When Claude’s safeguards blocked progress, the attacker switched to ChatGPT for evasion tactics and lateral movement strategies—creating a powerful “tag-team” cyber tool without needing advanced skills.
The haul was staggering: records on 195 million taxpayers from Mexico’s federal tax authority (SAT) and sensitive voter data from the National Electoral Institute (INE). Gambit researchers noted the breach’s scale democratizes cybercrime, turning AI into a force multiplier for novices.
Anthropic responded swiftly, banning the accounts and updating Claude Opus 4.6 with stronger protections. Mexican authorities launched federal probes months ago, though some local agencies deny breaches. The incident, first reported by Bloomberg, raises alarms about AI’s dual role in productivity and peril.